The element size problem

[shuaicai]

When the element size is set as 0.01m, the result worked out is not correct, unless the element size set as 1m, the the result worked out is correct and better, May I ask how is this going? Thank you! Cheng Xuelei.

[skamalzare]

Shuaicai,

I might be able to help you, please try to explain the problem a little more. What hits my mind now is that when you reduce your element size, you essentially will have more numerical computations. If your analysis setting applies numerical damping to your system (e.g. Newmark integration), you might affect your results with improper damping. To avoid this, you need to increase the accuracy of your solution. For example if you are using NormDispIncr to test the convergence, you might want to make it tighter.

Bests,
Soheil

[shuaicai]

skamalzare wrote:
> Shuaicai,
>
> I might be able to help you, please try to explain the problem a little
> more. What hits my mind now is that when you reduce your element size, you
> essentially will have more numerical computations. If your analysis setting
> applies numerical damping to your system (e.g. Newmark integration), you
> might affect your results with improper damping. To avoid this, you need to
> increase the accuracy of your solution. For example if you are using
> NormDispIncr to test the convergence, you might want to make it tighter.
>
> Bests,
> Soheil




The node number in model between the larger and smaller size is equal, as you said, what is value set in Newmark integration considering numerical damping? or HHT is used to avoid this problem?

[skamalzare]

You'll still have numerical damping with HHT. Numerical damping affects the simulation whenever your solution system performs integration. If you check the integration methods in Wiki, you'll see when the damping is introduced.

Generally, I use Newmark with gamma = 1.5 and Beta = 1.0 for gravity analyses to damp out oscillations, and gamma = 0.6 and beta = 0.3025 for dynamic analyses.

Don't forget, you should always have a proper convergence test limit to make sure, you are not getting affected by numerical errors. For example try using tighter convergence tests and see if your results will change. Say check:

test NormDispIncr 1e-3
test NormDispIncr 1e-5
test NormDispIncr 1e-7

Your results should stay the same, especially for the last two tolerances.

Bests,
Soheil

[shuaicai]

skamalzare wrote:
> You'll still have numerical damping with HHT. Numerical damping affects the
> simulation whenever your solution system performs integration. If you check
> the integration methods in Wiki, you'll see when the damping is introduced.
>
> Generally, I use Newmark with gamma = 1.5 and Beta = 1.0 for gravity
> analyses to damp out oscillations, and gamma = 0.6 and beta = 0.3025 for
> dynamic analyses.
>
> Don't forget, you should always have a proper convergence test limit to
> make sure, you are not getting affected by numerical errors. For example
> try using tighter convergence tests and see if your results will change.
> Say check:
>
> test NormDispIncr 1e-3
> test NormDispIncr 1e-5
> test NormDispIncr 1e-7
>
> Your results should stay the same, especially for the last two tolerances.
>
> Bests,
> Soheil


Hi Soheil, what's value about the thickness of model in plane strain? 1m was set in Wiki usually. What is the difference between 1m and 8m about the thickness? In my opinion, the length of single frame of underground structure is 8 m, and if the thickness set for 8m，the result would be more reasonable. And In my experience, the larger thincness value , easier computing convergence, but if I set larger thickness value than 1m in Wiki, Is there anything else need to be aware of?

[skamalzare]

Shuaicai,

Typically, the out-of-plane thickness is set very large to enforce plane strain conditions. I usually check different thicknesses to see which one gives me a better 2D condition. However, I have to say that for a real condition 10 m thickness, we compared results from FE models with 1 m, 10 m, 50 m, 100 m and 1000 m thicknesses. The results were almost the same. Therefore, we continued with 1 m thickness.

In another study at UC Davis, Chang et al. (2013) pretty much did the same and they also concluded that very large thicknesses resulted in negligible differences. Look at: FEM Analysis of Dynamic Soil-Pile-Structure Interaction in Liquefied and Laterally Spreading Ground.

Bests,
Soheil